bug fix

library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt

@@ -100,16 +100,18 @@ list(APPEND GEMM_INSTANCES
     device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp
     device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp)
 
-list(APPEND GEMM_INSTANCES
-    device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_default_instance.cpp
-    device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_default_instance.cpp
-    device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_default_instance.cpp
-    device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_padded_instance.cpp
-    device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_padded_instance.cpp
-    device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_padded_instance.cpp
-    device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_nk_mn_instance.cpp
-    device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_kn_mn_instance.cpp
-    device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_nk_mn_instance.cpp)
+if(GPU_TARGETS MATCHES "gfx94")
+  list(APPEND GEMM_INSTANCES
+      device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_default_instance.cpp
+      device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_default_instance.cpp
+      device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_default_instance.cpp
+      device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_padded_instance.cpp
+      device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_padded_instance.cpp
+      device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_padded_instance.cpp
+      device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_nk_mn_instance.cpp
+      device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_kn_mn_instance.cpp
+      device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_nk_mn_instance.cpp)
+endif()
 
 list(APPEND GEMM_INSTANCES
     device_gemm_wmma_f16_f16_f16_mk_kn_mn_instance.cpp

profiler/src/profile_grouped_gemm_fixed_nk.cpp

@@ -85,9 +85,11 @@ int profile_grouped_gemm_fixed_nk(int argc, char* argv[])
     const auto StrideCs = argToIntArray(argv[13]);
     const int kbatch    = argc == 15 ? std::stoi(argv[14]) : 1;
 
-    using F32  = float;
-    using F16  = ck::half_t;
-    using F8   = ck::f8_t;
+    using F32 = float;
+    using F16 = ck::half_t;
+#if defined(CK_ENABLE_FP8)
+    using F8 = ck::f8_t;
+#endif
     using BF16 = ck::bhalf_t;
     using I8   = int8_t;